How AI research and peer review could reinforce the wrong incentives

A problem of incentives, not identities

A review system can keep processing papers and rewarding researchers while becoming less effective at supporting discovery. That is the conditional warning in a preprint by Hongze Liu, Mengru Wang, Zhiyuan Zhang and Yao Liu, posted to arXiv on November 7, 2026. The supplied record identifies it as a working manuscript, with no journal publication listed.

The authors ask whether acceptance-oriented research and inexpensive, superficial evaluation could reinforce one another. Their contribution is a mathematical model supported by illustrative simulations—not a measurement showing that a particular scientific community has entered this state.

Crucially, the distinction is not between AI and human researchers. Either can pursue exploratory work or optimize for acceptance. Likewise, AI can assist consequential checking, while humans can rely on superficial signals.

Two populations adapting together

The researchers construct a two-population evolutionary game. Here, “evolutionary” means that practices earning higher private rewards increase their share of activity over time; it does not describe biological evolution.

Authors choose between acceptance-oriented research, organized around expected evaluation rewards, and exploratory research, which accepts greater uncertainty about recognition. Reviewers choose between proxy assessment, relying on inexpensive indicators of merit, and careful assessment, investing effort in understanding and checking important claims.

Authors’ rewards depend on acceptance benefits and research costs. Reviewers’ incentives depend on effort savings, workload and the privately valued benefit of accurate assessment.

The proposed feedback has two links. If acceptance-oriented production increases review pressure, proxy assessment becomes more attractive. If proxy assessment improves the relative rewards of acceptance-oriented work, authors have less incentive to explore. Neither link is assumed to hold universally.

Under specified conditions, the equations support two stable outcomes under identical institutional rules: exploration with careful review, or acceptance-oriented production with proxy review. An analytically derived “basin boundary” separates starting combinations that approach each outcome. In plain language, where the community begins can determine where it ends up.

Publication counts can conceal a loss

Stability alone does not establish scientific harm. The authors therefore introduce scientific value separately from acceptance rewards. This represents independently assessed contribution over a declared time horizon, rather than a review score or publication count.

In their illustrative baseline, both stable outcomes accept 0.65 papers per submission opportunity. But assumed contribution values produce accepted scientific value of 0.65 in the exploratory–careful outcome and 0.39 in the acceptance-oriented–proxy outcome: a modeled 40% reduction despite identical acceptance frequency.

These are normalized simulation values, not observed losses in science. Their dependence on assumptions is explicit: raising the assumed value of acceptance-oriented work makes the proxy outcome scientifically better without changing the behavioral trajectories. Acceptance-oriented work is not defined as inferior.

Faster reviewer adaptation also has no universally beneficial effect. It can spread careful assessment when exploratory authors remain common, or accelerate proxy assessment when acceptance-oriented production predominates.

What could interrupt the feedback?

The paper analyzes two interventions. Externally resourced careful assessments change the evaluation authors encounter; rewards for careful reviewing change reviewers’ incentives.

In the baseline example, assigning careful assessments to more than 75% of opportunities, or providing a careful-review reward above 0.2 normalized payoff units, is sufficient for eventual recovery from starting states containing both strategies. These are model-specific thresholds, not recommended institutional targets. Assigned assessments must determine outcomes and require externally supplied capacity.

Weaker support can help some starting communities while leaving an established proxy regime intact. Temporary support can also produce lasting change if it moves the community across the baseline basin boundary before withdrawal. Model time has no conversion into months or conference cycles.

The simulations report gross scientific value, not benefits after subtracting verification resources or financing costs. They do not identify an optimal policy.

A conditional warning, not a diagnosis

The baseline compresses diverse practices into binary choices, simplifies acceptance decisions and omits strategic foresight and shocks. Alternative specifications sometimes preserve its contrasting outcomes, but others produce different mixtures or continuing cycles.

The cited empirical literature motivates possible vulnerabilities; it does not calibrate the model’s feedback coefficients. Testing the mechanism requires jointly measuring evaluation rewards, workload-driven checking behavior and independently assessed contribution. Tool-use percentages alone cannot establish it.

The authors also propose institutional redesigns, including auditable AI committees with human oversight and rewards for claim-level verification. These remain proposals, not demonstrated results.

This account draws on the supplied full manuscript extraction, which may omit visual information or damage equations. Its central lesson is conditional but clear: accelerating production and scoring is not equivalent to making validated discovery and careful checking worthwhile.